Adiabatic time-dependent density functional methods for excited state

properties
Filipp Furche and Reinhart Ahlrichs

Citation: J. Chem. Phys. 117, 7433 (2002); doi: 10.1063/1.1508368

View online: http://dx.doi.org/10.1063/1.1508368
View Table of Contents: http://jcp.aip.org/resource/1/JCPSA6/v117/i16
Published by the American Institute of Physics.

Additional information on J. Chem. Phys.

Journal Homepage: http://jcp.aip.org/
Journal Information: http://jcp.aip.org/about/about_the_journal
Top downloads: http://jcp.aip.org/features/most_downloaded
Information for Authors: http://jcp.aip.org/authors

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 JOURNAL OF CHEMICAL PHYSICS VOLUME 117, NUMBER 16 22 OCTOBER 2002

Adiabatic time-dependent density functional methods

for excited state properties
Filipp Furchea) and Reinhart Ahlrichs
Institut für Physikalische Chemie, Universität Karlsruhe, Kaiserstraße 12, 76128 Karlsruhe, Germany
共Received 22 May 2002; accepted 30 July 2002兲
This work presents theory, implementation, and validation of excited state properties obtained from
time-dependent density functional theory 共TDDFT兲. Based on a fully variational expression for the
excited state energy, a compact derivation of first order properties is given. We report an
implementation of analytic excited state gradients and charge moments for local, gradient corrected,
and hybrid functionals, as well as for the configuration interaction singles 共CIS兲 and time-dependent
Hartree–Fock 共TDHF兲 methods. By exploiting analogies to ground state energy and gradient
calculations, efficient techniques can be transferred to excited state methods. Benchmark results
demonstrate that, for low-lying excited states, geometry optimizations are not substantially more
expensive than for the ground state, independent of the molecular size. We assess the quality of
calculated adiabatic excitation energies, structures, dipole moments, and vibrational frequencies by
comparison with accurate experimental data for a variety of excited states and molecules. Similar
trends are observed for adiabatic excitation energies as for vertical ones. TDDFT is more robust than
CIS and TDHF, in particular, for geometries differing significantly from the ground state minimum.
The TDDFT excited state structures, dipole moments, and vibrational frequencies are of a
remarkably high quality, which is comparable to that obtained in ground state density functional
calculations. Thus, yielding considerably more accurate results at similar computational cost,
TDDFT rivals CIS as a standard method for calculating excited state properties in larger
molecules. © 2002 American Institute of Physics. 关DOI: 10.1063/1.1508368兴

I. INTRODUCTION applications.7 In the past five years, time-dependent density

There is a broad interest in properties of electronic ex-
cited states ranging from spectroscopy to photochemistry and
biology.1,2 In fact, for a deeper understanding of most pho-
tophysical and photochemical phenomena, some information
on excited state potential energy surfaces is necessary. For
example, luminescence and related emission processes are
associated with geometric relaxation of excited states. It is
also well known that structural changes upon electronic ex-
citation can significantly alter the shape of absorption and
circular dichroism spectra. According to the Franck–Condon
principle, knowledge of the excited state normal modes is
required to describe the fine structure observed in high-
resolution spectra. Vibrations may induce forbidden transi-
tions, or mediate nonadiabatic coupling. Information on
structure and dynamics of excited states is the most impor-
tant prerequisite for modeling photochemical reactions.3
With the development of new techniques in laser
spectroscopy,4 excited state properties, e.g., vibrational fre-
quencies or dipole moments, can be probed selectively for
increasingly large systems. Reliable theoretical models are
necessary, however, to explain the experimental data in terms
of the excited state geometric and electronic structure. Up to
now, the standard for calculating excited state equilibrium
properties of larger molecules is the configuration interaction
singles 共CIS兲 method.5,6 However, due to a neglect of elec-
tron correlation, CIS results are not accurate enough in many
a兲
Electronic mail: filipp.furche@chemie.uni-karlsruhe.de
functional theory 共TDDFT兲8 has almost completely super-
seded CIS in vertical excitation energy calculations, yielding
considerably improved results at similar computational
cost.9–11 The pioneering work of Amos and Van Caillie12,13
and recent case studies14,15 indicate that TDDFT may be
more accurate than CIS for excited state potential energy
surfaces and other properties as well.
A number of state-based density functional methods has
been proposed for excited state treatments;16 –18 for the low-
est excited singlet state, a spin-restricted open-shell Kohn–
Sham 共KS兲 type scheme has been proposed.19 It largely re-
mains to be shown that these methods are competitive to
TDDFT in standard applications. Judging from experience
with wave-function methods,20 advantages of TDDFT, such
as a consistent definition of transition moments and the ex-
plicit treatment of near-degeneracy effects between singly
excited configurations, will not be easily overcome.
The purpose of the present work is threefold. First, we
give a TDDFT excited state energy functional which is fully
variational with respect to all parameters. This considerably
simplifies the derivation of analytic expressions for first or-
der excited state properties. We extend the theory for general
spin-unrestricted ground states, recovering previous results
for CIS,5 time-dependent Hartree–Fock 共TDHF兲,21 and
TDDFT13 as special cases. Second, we describe an efficient
implementation of excited state gradients with respect to
nuclear coordinates and other first order properties for the
above methods. In contrast to numerical differentiation, the

0021-9606/2002/117(16)/7433/15/$19.00 7433 © 2002 American Institute of Physics

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 7434 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002
effort for calculating gradients analytically is independent of
the number of nuclear degrees of freedom. We emphasize the
analogy to ground state energy and gradient calculations,
which allows us to transfer proven techniques to excited state
methods. The feasibility of excited state geometry optimiza-
tions is investigated in typical applications. Finally, we as-
sess the quality of calculated adiabatic excitation energies,
excited state dipole moments, structures, and vibrational fre-
quencies by comparison with accurate experimental data.
Our test sample contains a variety of states of different char-
acter and spin multiplicity. The methods include, besides CIS
and TDHF, the most popular functionals used in TDDFT
calculations. We analyze the results and discuss implications
for theoretical treatments of excited states.
II. THEORY
A. Excited state properties in TDDFT
The variational formulation of TDDFT 共Ref. 22兲 pro-
vides a very convenient framework for the calculation of
excited state properties. The basic statement is that excitation
energies are the stationary points of the functional
G 关 X,Y ,⍀ 兴 ⫽ 具 X,Y 兩 ⌳ 兩 X,Y 典 ⫺⍀ 共 具 X,Y 兩 ⌬ 兩 X,Y 典 ⫺1 兲 . 共1兲
⍀ is a real Lagrange multiplier, and the vectors
兩 X,Y 典 ⫽ 冉冊X
Y
are defined on the Hilbert space L⫽L virt⫻L occ 丣 L occ⫻L virt ,
where L occ and L virt denote the Hilbert spaces of occupied
共occ兲 and virtual 共virt兲 molecular orbitals 共MOs兲, respec-
tively. The MOs ␸ p ␴ (r) are solutions of the static KS equa-
tions with eigenvalues ⑀ p ␴ . As usual, indices i, j,... label
occupied, a,b,... virtual, and p,q,... general MOs. We as-
sume the MOs to be real, which is always possible in the
absence of magnetic fields. The ‘‘superoperators’’ ⌳ and ⌬
read
⌳⫽ 冉 冊
A
B
B
A
, ⌬⫽ 冉 1
0
0
⫺1
冊 .
A and B are sometimes called orbital rotation Hessians and
have the matrix representation
共 A⫹B 兲 ia ␴ jb ␴ ⬘ ⫽ 共 ⑀ a ␴ ⫺ ⑀ i ␴ 兲 ␦ i j ␦ ab ␦ ␴␴ ⬘
⫹2 共 ia ␴ 兩 jb ␴ ⬘ 兲 ⫹2 f ia ␴ jb ␴ ⬘
xc
⫺c x␦ ␴␴ ⬘ 关共 ja ␴ 兩 ib ␴ 兲 ⫹ 共 ab ␴ 兩 i j ␴ 兲兴 ,
共 A⫺B 兲 ia ␴ jb ␴ ⬘ ⫽ 共 ⑀ a ␴ ⫺ ⑀ i ␴ 兲 ␦ i j ␦ ab ␦ ␴␴ ⬘ ⫹c x␦ ␴␴ ⬘ 关共 ja ␴ 兩 ib ␴ 兲
⫺ 共 ab ␴ 兩 i j ␴ 兲兴 .
( pq ␴ 兩 rs ␴ ⬘ ) is a two-electron repulsion integral in Mulliken
xc
notation, and f pq ␴ rs ␴ ⬘ represents a matrix element of the
exchange-correlation kernel in the adiabatic approximation,
␦ 2 E xc
f ␴␴ ⬘ 共 r,r⬘ 兲 ⫽
xc
. 共5兲
␦ ␳ ␴ 共 r兲 ␦ ␳ ␴ ⬘ 共 r⬘ 兲
E xc denotes the static exchange-correlation energy func-
tional; the functional derivative is evaluated, as all other
Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
F. Furche and R. Ahlrichs
functionals in the present work, at the static ground state
density. The hybrid mixing parameter c x introduced by
Becke23,24 allows us to interpolate between the limits of
‘‘pure’’ density functionals (c x ⫽0) and TDHF theory (c x
⫽1, E xc⫽0), which is also known as the random phase
approximation 共RPA兲 for excitation energies.25 The CIS or
Tamm–Dancoff approximation 共TDA兲 is easily derived by
constraining Y identically to zero in the variation of G. A
TDA type scheme has recently been proposed in a TDDFT
context as well.11 All these methods are contained as special
cases in the present approach.
Applying the variational principle, we obtain the station-
arity conditions for G,
␦G
⫽ 共 ⌳⫺⍀⌬ 兲 兩 X,Y 典 ⫽0, 共6a兲
␦ 具 X,Y 兩
⳵G
⫽ 具 X,Y 兩 ⌬ 兩 X,Y 典 ⫺1⫽0. 共6b兲
⳵⍀
Equation 共6a兲 is the well-known time-dependent Kohn-Sham
共TDKS兲 eigenvalue problem; solution under constraint 共6b兲
yields the TDDFT excitation energies as eigenvalues ⍀ and
TDKS transition density matrices as eigenvectors 兩 X,Y 典 . Un-
共2兲
less stated otherwise, we shall assume in the following that
Eqs. 共6a兲 and 共6b兲 have been solved for a particular excited
state under consideration.
In analogy to ground state properties, excited state prop-
erties are conveniently defined as derivatives of the excited
state energy with respect to an external perturbation.26 These
derivatives are, somewhat symbolically, denoted by a super-
script ␰. We think of ␰ as a nuclear coordinate in the first
place, leading to excited state energy gradients with respect
to the positions of the nuclei. However, other excited state
properties are accessible as well. If ␰ represents, e.g., a com-
ponent of a static electric field, excited state dipole moments
共3兲
can be computed. The total excited state energy is the sum of
the ground state energy and the excitation energy. Thus, ex-
cited state properties are a sum of the corresponding ground
state properties and an excitation part given by derivatives of
the excitation energy. We can focus on the latter here since
the computation of HF and KS ground state properties is
certainly well established. In terms of the functional G de-
fined in Eq. 共1兲,
共4a兲
⍀ ␰ ⫽G ␰ 关 X,Y ,⍀ 兴 ⫽ 具 X,Y 兩 ⌳ ␰ 兩 X,Y 典 . 共7兲
共4b兲 It is important to note that due to the variational principle,
first order properties do not require derivatives of the excita-
tion vectors 兩 X,Y 典 . If f denotes the number of nuclear de-
grees of freedom in a molecule, computing first derivatives
of 兩 X,Y 典 would be roughly f times as expensive as a single-
point excitation energy calculation.
However, a straightforward calculation of ⌳ ␰ still in-
volves derivatives of the MO coefficients. In basis set meth-
ods, the MOs are usually expanded as a linear combination
of atomic orbitals 共LCAO兲,
 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002 TDDFT for excited state properties 7435

The Lagrangian L is required to be stationary with respect to

␸ p ␴ 共 r兲 ⫽ 兺␮ C ␮ p ␴ ␹ ␮共 r兲 .
A MO ␸ p ␴ can depend on the perturbation through the
atomic orbitals 共AOs兲 ␹ ␮ (r) and the expansion coefficients
C. Derivatives of the basis functions ␹ ␮ (r) are easily dealt
with: In case ␰ denotes a nuclear coordinate, the derivatives
␹ ␮␰ (r) are nonzero only if ␰ refers to the atom ␹ ␮ (r) is cen-
tered at. The MO coefficients, though, depend on all f per-
turbations in general, so that a calculation of the derivatives
C ␰ would require f times the work of a calculation of the
unperturbed MOs. Fortunately, it is possible to eliminate de-
rivatives C ␰ by introducing so-called relaxed densities, at a
computational cost independent of f . This is a consequence
of the Sternheimer–Dalgarno interchange theorem,27,28
which is better known as the Z vector method29 in quantum
chemistry. However, the introduction of perturbed MO coef-
ficients leads to lengthy expressions that have to be rear-
ranged in a nontrivial way in order to apply the Z vector
method. Although tedious, this procedure has previously
been used to derive excited state gradients for CIS,5 TDHF,21
and TDDFT12,13 methods.
共8兲 all its parameters, X, Y , ⍀, C, Z, and W, which are consid-
ered as variables in Eq. 共12兲. The sum of the ground state
Lagrangian and L thus represents a fully variational expres-
sion for the excited state energy functional. Similar tech-
niques have been used in correlated treatments of excited
state properties.31 While variation of L with respect to 兩 X,Y 典
and ⍀ leads back to Eqs. 共6兲, the additional Lagrange mul-
tipliers Z and W enforce the conditions
⳵L
⫽F ia ␴ ⫽0, 共13兲
⳵ Z ia ␴
⳵L
⫽S pq ␴ ⫺ ␦ pq ⫽0. 共14兲
⳵ W pq ␴
The MOs ␸ p ␴ (r) are thus constrained to satisfy the static KS
equations 共up to unitary equivalence兲, and to be orthonormal.
These conditions fix the MO coefficients C for all values of
the external perturbation. Z and W themselves are deter-
mined from the condition

⳵L
B. Lagrangian of the excitation energy ⫽0, 共15兲
⳵ C ␮p␴
We pursue a different strategy which avoids MO coeffi-
cient derivatives from the very beginning instead of first in- as shown in detail below. Solving Eq. 共15兲 is approximately
troducing them and reeliminating them later. In a first step, f times more efficient than computing perturbed MO coeffi-
we replace the diagonal part of (A⫹B) and (A⫺B) in Eqs. cients.
共4a兲 and 共4b兲 by At first sight, the definition of excitation energies as sta-
tionary points of L appears to be more complicated than the
共 F ab ␴ ␦ i j ⫺F i j ␴ ␦ ab 兲 ␦ ␴␴ ⬘ . 共9兲 definition as stationary points of G, because additional pa-
The effective KS one-particle Hamiltonian or Fock operator rameters have been introduced. However, while G depends
implicitly through the MO coefficients on an external pertur-
F pq ␴ ⫽h pq ␴ ⫹ 兺 关共 pq ␴ 兩 ii ␴ ⬘ 兲 ⫺c x␦ ␴␴ ⬘ 共 pi ␴ 兩 iq ␴ 兲兴 bation, this implicit dependence drops out of L by virtue of
i␴ ⬘ the variational principle. Thus, L is an explicit functional of
any external perturbation. This is a crucial advantage in the
⫹V xc
pq ␴ , 共10兲
treatment of excited state properties: Once X, Y , ⍀, C, Z,
consists of the core Hamilton h, a Coulomb and 共hybrid-兲 and W have been determined from the stationarity condi-
exchange part, and the exchange-correlation potential tions, derivatives of the excitation energy follow as
␦ E xc
V ␴xc共 r兲 ⫽ . 共11兲 ⍀ ␰ ⫽L ␰ 关 X,Y ,⍀,C,Z,W 兴
␦ % ␴ 共 r兲
In the basis of canonical KS MOs, F is diagonal with the
eigenvalues ⑀ p ␴ as diagonal entries. This choice is somewhat
⫽G ( ␰ ) 关 X,Y ,⍀ 兴 ⫹ 兺
ia ␴
(␰)
Z ia ␴ F ia ␴⫺ 兺
pq ␴ ,p⭐q
W pq ␴ S (pq␰ )␴ .

arbitrary, though, since all physical properties are invariant 共16兲

under unitary transformations of occupied and virtual orbit-
als among themselves. The above replacement renders the In contrast to Eq. 共7兲, the derivatives are straightforward to
excited state formalism manifestly invariant under these evaluate, because the MO coefficients C are held constant at
transformations. The constraint of a canonical choice of the their zero order values, as indicated by the superscript 共␰兲.
MOs can thus be dropped when calculating derivatives of the This is especially obvious for the contribution
excitation energy.
In a slight variation of an idea of Helgaker and G ( ␰ ) 关 X,Y ,⍀ 兴
Jørgensen,30 we define the auxiliary functional
⫽ 具 X,Y 兩 ⌳ ( ␰ ) 兩 X,Y 典
L 关 X,Y ,⍀,C,Z,W 兴 ⫽G 关 X,Y ,⍀ 兴 ⫹ 兺 Z ia ␴ F ia ␴ 1
ia ␴ ⫽ 兺
2 ia ␴ jb ␴ ⬘
(␰)
关共 A⫹B 兲 ia ␴ jb ␴ ⬘ 共 X⫹Y 兲 ia ␴ 共 X⫹Y 兲 jb ␴ ⬘

⫺ 兺
pq ␴ ,p⭐q
W pq ␴ 共 S pq ␴ ⫺ ␦ pq 兲 . 共12兲 (␰)
⫹ 共 A⫺B 兲 ia ␴ jb ␴ ⬘ 共 X⫺Y 兲 ia ␴ 共 X⫺Y 兲 jb ␴ ⬘ 兴 . 共17兲

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 7436 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002 F. Furche and R. Ahlrichs

C. Determination of the Lagrange multipliers Z and W where the occ–occ and virt–virt parts of Z are zero. The
physical meaning of P will become obvious in the next sub-
The orbital stationarity constraint 共15兲 gives rise to four
section. The remaining conditions from Eq. 共15兲 fix the
conditions for Z and the various parts of W. A detailed deri-
Lagrange multipliers W according to
vation is given in Appendix A; we merely state the results
here. For Z one obtains
W i j␴⫽ 兺a ⍀ 兵 共 X⫹Y 兲 ia ␴共 X⫺Y 兲 ja ␴ ⫹ 共 X⫺Y 兲 ia ␴
兺 共 A⫹B 兲 ia ␴ jb ␴ ⬘ Z jb ␴ ⬘ ⫽⫺R ia ␴ , 共18兲
jb ␴⬘
which is traditionally called the Z vector equation. Defining ⫻ 共 X⫹Y 兲 ja ␴ 其 ⫺ 兺a ⑀ a ␴ 兵 共 X⫹Y 兲 ia ␴共 X⫹Y 兲 ja ␴
the unrelaxed difference density matrix T
⫹ 共 X⫺Y 兲 ia ␴ 共 X⫺Y 兲 ja ␴ 其 ⫹H ⫹
i j␴关 P 兴
T ab ␴ ⫽ 1
2 兺i 兵 共 X⫹Y 兲 ia ␴共 X⫹Y 兲 ib ␴
⫹2 兺
kc ␴ ⬘ ld ␴ ⬙
xc
g i j ␴ kc ␴ ⬘ ld ␴ ⬙ 共 X⫹Y 兲 kc ␴ ⬘ 共 X⫹Y 兲 ld ␴ ⬙ ,
⫹ 共 X⫺Y 兲 ia ␴ 共 X⫺Y 兲 ib ␴ 其 ,
共19兲
T i j ␴ ⫽⫺ 1
2 兺a 兵 共 X⫹Y 兲 ia ␴共 X⫹Y 兲 ja ␴ W ab ␴ ⫽ 兺i ⍀ 兵 共 X⫹Y 兲 ia ␴共 X⫺Y 兲 ib ␴
⫹ 共 X⫺Y 兲 ia ␴ 共 X⫺Y 兲 ja ␴ 其 , ⫹ 共 X⫺Y 兲 ia ␴ 共 X⫹Y 兲 ib ␴ 其
T ia ␴ ⫽T ai ␴ ⫽0,
⫹ 兺i ⑀ i ␴ 兵 共 X⫹Y 兲 ia ␴共 X⫹Y 兲 ib ␴
and, for arbitrary vectors V pq ␴ , the linear transformations
⫹ 共 X⫺Y 兲 ia ␴ 共 X⫺Y 兲 ib ␴ 其 , 共24兲
H⫹
pq ␴ 关 V 兴 ⫽ 兺 兵 2 共 pq ␴ 兩 rs ␴ ⬘ 兲 ⫹2 f pq ␴ rs ␴ ⬘ ⫺c x␦ ␴␴ ⬘
xc

rs ␴ ⬘
⫻ 关共 ps ␴ 兩 rq ␴ 兲 ⫹ 共 pr ␴ 兩 sq ␴ 兲兴 其 V rs ␴ , 共20a兲 W ia ␴ ⫽ 兺j 兵 共 X⫹Y 兲 ja ␴ H ⫹ji ␴ 关 X⫹Y 兴
⫹ 共 X⫺Y 兲 ia ␴ H ⫺
ji ␴ 关 X⫺Y 兴 其 ⫹ ⑀ i ␴ Z ia ␴ .
H⫺
pq ␴ 关 V 兴 ⫽ 兺 c x␦ ␴␴ ⬘ 关共 ps ␴ 兩 rq ␴ 兲 ⫺ 共 pr ␴ 兩 sq ␴ 兲兴 V rs ␴ ,
rs ␴ ⬘
共20b兲 W will turn out as the energy-weighted difference density
the right-hand side 共RHS兲 R is conveniently expressed as matrix below.

R ia ␴ ⫽ 兺b 兵 共 X⫹Y 兲 ib ␴ H ab⫹ ␴ 关 X⫹Y 兴

D. First order properties of excited states
⫺
⫹ 共 X⫺Y 兲 ib ␴ H ab ␴ 关 X⫺Y 兴 其 Using the definitions of A, B, and F, Eq. 共16兲 is easily
transformed into the final form
⫺ 兺j 兵 共 X⫹Y 兲 ja ␴ H ⫹ji ␴ 关 X⫹Y 兴
⫹ 共 X⫺Y 兲 ia ␴ H ⫺ ⫹
ji ␴ 关 X⫺Y 兴 其 ⫹H ia ␴ 关 T 兴
⍀ ␰⫽ 兺
␮ ␯␴
h ␮␰ ␯ P ␮ ␯␴ ⫺ 兺
␮ ␯␴
S ␮␰ ␯ W ␮ ␯␴ ⫹ 兺
␮ ␯␴
(␰)
V ␮xc␯␴ P ␮ ␯␴

⫹2 兺
jb ␴ ⬘ kc ␴ ⬙
xc
g ia ␴ jb ␴ ⬘ kc ␴ ⬙ 共 X⫹Y 兲 jb ␴ ⬘ 共 X⫹Y 兲 kc ␴ ⬙ . ⫹ 兺 共 ␮ ␯ 兩 ␬ ␭ 兲 ␰ ⌫ ␮ ␯␴ ␬ ␭ ␴ ⬘
␮ ␯ ␬ ␭ ␴␴ ⬘
共21兲
xc
g pq ␴ rs ␴ ⬘ tu ␴ ⬙ denotes a matrix element of the third order func-
⫹ 兺 xc ( ␰ )
f ␮ ␯␴ ␬ ␭ ␴ ⬘ 共 X⫹Y 兲 ␮ ␯␴ 共 X⫹Y 兲 ␬ ␭ ␴ ⬘ . 共25兲
␮ ␯ ␬ ␭ ␴␴ ⬘
tional derivative
␦ 3 E xc Greek indices indicate that operator matrix elements are
g ␴␴ ⬘ ␴ ⬙ 共 r,r⬘ ,r⬙ 兲 ⫽
xc
. 共22兲 taken in the AO basis, and
␦ ␳ ␴ 共 r兲 ␦ ␳ ␴ ⬘ 共 r⬘ 兲 ␦ ␳ ␴ ⬙ 共 r⬙ 兲
From Eq. 共21兲 it appears that the computation of R is rather
involved; however, as discussed in Sec. III, it is never nec- V ␮ ␯␴ ⫽ 兺
pq
C ␮ p ␴ V pq ␴ C ␯ q ␴ , 共26兲
essary to set up and store more than two-index quantities.
After the Z vector Eq. 共18兲 has been solved, the relaxed
one-particle difference density matrix P is obtained as for V⫽ P, W, and (X⫾Y ). If D denotes the ground state
density matrix, the effective two-particle difference density
P⫽T⫹Z, 共23兲 matrix is given by

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002
⌫ ␮ ␯␴ ␬ ␭ ␴ ⬘ ⫽ 21 兵 2 P ␮ ␯␴ D ␬ ␭ ␴ ⬘ ⫹2 共 X⫹Y 兲 ␮ ␯␴ 共 X⫹Y 兲 ␬ ␭ ␴ ⬘
⫺c x␦ ␴␴ ⬘ 关 P ␮ ␭ ␴ D ␬ ␯␴ ⬘ ⫹ P ␮␬ ␴ D ␭ ␯␴ ⬘
⫹ 共 X⫹Y 兲 ␮ ␭ ␴ 共 X⫹Y 兲 ␬ ␯␴ ⬘
⫹ 共 X⫹Y 兲 ␮␬ ␴ 共 X⫹Y 兲 ␭ ␯␴ ⬘
⫺ 共 X⫺Y 兲 ␮ ␭ ␴ 共 X⫺Y 兲 ␬ ␯␴ ⬘
⫹ 共 X⫺Y 兲 ␮␬ ␴ 共 X⫺Y 兲 ␭ ␯␴ ⬘ 兴 其 ;
the various direct and exchange contributions thus separate
into products of lower rank quantities, which is important for
the processing of ⌫.
Equation 共25兲 indicates that, after spin summation, P,
W, and ⌫ may be regarded as partial derivatives of the exci-
tation energy with respect to h ␮ ␯ , S ␮ ␯ , and the electron–
electron interaction. All first order properties can be ex-
pressed in the form of Eq. 共25兲 since the excitation energy
can depend on an external perturbation only through h, S,
the electron–electron interaction, and the exchange-
correlation functional derivatives. For example, the differ-
ence between the excited and ground state dipole moments
follows as
⌬ 具 ␮典 ⫽tr共 P ␮兲 ,
if ␮ denotes the dipole moment operator. Population analysis
of P can elucidate the charge rearrangement and the change
in bond order induced by an electronic excitation. This is
complementary to the information supplied by the vectors
兩 X,Y 典 , which describe the excitation in terms of occupied
and virtual MOs, i.e., the change in the electronic structure.
The Z vector contribution to P in Eq. 共23兲, which accounts
for orbital relaxation effects, is often of the same order of
magnitude as the ‘‘unrelaxed’’ part T and cannot be ne-
glected, in contrast to what has occasionally been assumed.32
We note in passing that, in analogy to the ground state
KS scheme, P would yield the exact density difference if the
exact 共time-dependent兲 exchange-correlation functional were
used. This follows from the fact that the density computed
from P is identical to the functional derivative of the excita-
tion energy with respect to a local external potential.
III. IMPLEMENTATION
A. General strategy
The results of the last section suggest that the analytic
computation of excited state energy gradients and other prop-
erties should be organized as follows: Given the solution of
the ground state KS equations,
共1兲 determine the excitation energy ⍀ and excitation vector
兩 X,Y 典 from Eqs. 共6兲;
共2兲 solve the Z vector Eq. 共18兲 to obtain the Lagrange mul-
tipliers Z and W; and
共3兲 evaluate the gradient and other properties according to
Eq. 共25兲.
Methods for solving the TDKS eigenvalue problem 共6a兲 and
the static coupled KS Eqs. 共18兲 are well developed.10,33–35 In
an iterative integral direct algorithm, the operations neces-
Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
TDDFT for excited state properties 7437
sary in each iteration closely resemble those for a single
ground state Fock matrix construction. If N is a measure of
the system size, the use of integral prescreening leads to an
asymptotic scaling of computational cost proportional to N 2
for the Coulomb and hybrid exchange contributions, while
共nearly兲 linear scaling of the exchange-correlation contribu-
tion is achieved by efficient molecular quadrature
schemes.36,37 The total effort for steps 共1兲 and 共2兲 is thus
共27兲 comparable to solving the ground state KS equations. For
higher excited states the first step becomes increasingly ex-
pensive, since all lower states of the same symmetry have to
be calculated as well. This is not a significant limitation,
however, since in the majority of applications highly excited
states do not play an important role.
Step 共3兲, evaluation of the gradient, is largely identical to
the analytic computation of ground state gradients. In fact,
apart from the exchange-correlation terms discussed below,
only the expressions for the one- and two-particle density
matrix as well as the energy-weighted density matrix have to
be replaced by P, W, and ⌫. By adding the ground state
contributions to P, W, and ⌫ before evaluating Eq. 共25兲,
gradients of the total excited state energy can be obtained in
a one-step procedure. The cost for step 共3兲 is thus virtually
共28兲 identical to that for ground state gradients.
To sum up, following the method outlined above, the
effort for an excited state geometry optimization is not sub-
stantially higher than for a ground state geometry optimiza-
tion. It is important to stress the similarity to the computation
of the ground state energy and gradient since this implies that
the whole arsenal of efficient techniques for ground states is
applicable to excited states, too. If there is progress in
ground state methods, excited state methods will also benefit.
Finally, given a code for TDDFT excitation energies and
response properties as well as a ground state gradient code,
excited state gradients can be implemented by minor modi-
fications.
B. Exchange-correlation contributions
The only terms which are not straightforward to transfer
from ground state and linear response calculations are those
containing third-order functional derivatives 关Eqs. 共21兲 and
共24兲兴 and geometric derivatives of the exchange-correlation
potential and kernel 关Eq. 共25兲兴. However, scrutiny of these
contributions reveals that they can be reduced to the same
form as V ␮xc␯␴ and E xc ( ␰ ) , both routinely computed in ground
state energy and gradient calculations. To put this into more
explicit terms, consider the exchange-correlation energy
functional within the generalized gradient approximation
共GGA兲,
E xc⫽ 冕 d 3 r f 共 ␳ ␣ , ␳ ␤ , ␥ ␣␣ , ␥ ␣␤ , ␥ ␤␤ 兲 , 共29兲
where ␥ ␴␴ ⬘ ⫽ⵜ ␳ ␴ ⵜ ␳ ␴ ⬘ . The local density approximation
共LDA兲, where f depends only on the spin densities ␳ ␣ and
␳ ␤ , is a special case of the GGA. Now, the matrix of the
exchange-correlation potential is computed by numerical
quadrature of
 7438 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002 F. Furche and R. Ahlrichs

TABLE I. Survey of the exchange-correlation energy functionals used in this work.

Name Type Exchange Correlation Ref.

SVWN LDA S VWN 52–54

PW LDA S PW 55
BLYP GGA S⫹B88 LYP 56, 57
BP86a GGA S⫹B88 VWN⫹P86 58
PBE GGA S⫹PBE共X兲 PW⫹PBE共C兲 59
PBE0 hybrid 0.25 HF⫹0.75 关 S⫹PBE共X兲兴 PW⫹PBE共C兲 60
B3LYPb hybrid 0.2 HF⫹0.72 B88⫹0.8 S 0.19 VWN⫹0.81 LYP 24
a
The BP86 implementation in TURBOMOLE and CADPAC differs from that in GAUSSIAN and Q-CHEM by the LDA
correlation part, which is VWN instead of PZ81 共Ref. 61兲.
b
At variance with the GAUSSIAN and Q-CHEM programs, we use the full 共beyond RPA兲 VWN parametrization in
the spin-polarized case 共also called VWN-V兲, as recommended by VWN.

V ␮xc␯␴ ⫽ 冕 冉 ⳵␳⳵
d 3r
f
␴
⫹2
⳵f
⳵ ␥ ␴␴
ⵜ ␳ ␴ⵜ
ground states. As is well known, in this case distinguishing
between singlet and triplet excitations further reduces CPU
and storage requirements by a factor of 2. Nevertheless,
⫹
⳵f
⳵ ␥ ␴␴ ⬘ ␴ ⬘ 冊
ⵜ ␳ ⵜ ␹ ␮ 共 r兲 ␹ ␯ 共 r兲 , 共30兲
EGRAD can also deal with general spin unrestricted open
shell references, thus permitting the treatment of excited
states of radicals. We have carefully checked the correctness
and the partial derivatives of E xc are calculated as of our implementation by comparison with results from nu-

E xc ( ␰ ) ⫽ 冕 d 3r 兺␴ 冉 ⳵f
⳵␳ ␴
⫹2
⳵f
⳵ ␥ ␴␴
ⵜ ␳ ␴ⵜ
merical differentiation.
In Table II we give some representative CPU timings for
the calculation of the excited state energy⫹gradient and

⫹
⳵f
⳵ ␥ ␴␴ ⬘ ␴ ⬘ 冊
ⵜ ␳ ⵜ ␳ ␴( ␰ ) 共 r兲 , 共31兲
compare them to timings for ground state energies. The num-
bers show that geometry optimizations for lower excited
states are not significantly more expensive than ground state
where ␴ ⬘ ⫽ ␴ , as is of course well known.36,38 The contribu- geometry optimizations. The asymptotic scaling of computa-
tions necessary for excited state gradients are obtained by tional cost is approximately proportional to N 2 , thus allow-
formally replacing first derivatives of f by second and third ing for rather large applications already. The time-
derivatives, multiplied by an effective density and its gradi- determining step, computation and processing of the two-
ents on the quadrature grid. The latter is given by electron repulsion integrals, can be further accelerated by
means of the resolution of the identity 共RI兲 technique;40,41
兺
␮␯
共 X⫹Y 兲 ␮ ␯␴ ␹ ␮ 共 r兲 ␹ ␯ 共 r兲 , 共32兲 work along these lines is in progress.

i.e., the first order density response, for the terms containing
second and third functional derivatives, and by IV. COMPUTATIONAL DETAILS

兺
␮␯
P ␮ ␯␴ ␹ ␮ 共 r兲 ␹ ␯ 共 r兲 , 共33兲 The structures of the molecules considered below were
optimized in the ground and excited states 共energy change
(␰)
i.e., the difference density, for the contribution from V ␮xc␯␴ in
Eq. 共25兲. The only additional operations required for excited
state gradient calculations are thus the evaluation of the ef- TABLE II. CPU timings 共hours兲 for single point ground state energies
fective densities and the second and third derivatives of f . 共DSCF兲 and single point excitation energies⫹excited state gradients 共EGRAD兲.
The calculations were performed on a single CPU of an HP J240 worksta-
Analytic implementation of the latter for the GGA function- tion 共240 MHz兲. Start orbitals were generated by extended Hückel theory,
als used in this work is tedious, but elementary. and a convergence of the ground state density matrix to 10⫺7 a.u. was
required. Quadrature grids were of size 3 共Ref. 36兲. N BF denotes the number
C. The program EGRAD
of contracted Gaussian basis functions. The basis sets are described in
Following the strategy outlined in Sec. III A, TDDFT Sec. IV.
excited state gradients have been implemented in the new System Sym. State Funct. Basis N BF DSCF EGRAD
module EGRAD, which is part of the TURBOMOLE program 1
trans-stilbene C 2h 1 Bu B3LYP SVP 270 0:13 0:23
suite.39 A survey of the most important functionals currently
Triphenylmethyla C3 1 2
E BP SVP 360 0:18 0:32
available and used in this work is given in Table I; in addi- 关6兴Helicene C2 1 3
A BP SV共P兲 422 1:22 1:17
tion, the CIS and TDHF methods are available. We fully DMABNb C 2v 2 1
A1 BP TZVPP 546 2:25 2:54
exploit molecular point group symmetry for all finite point Retinal C1 2 1
A PBE SVP 455 2:33 2:28
groups 共with less than 100-fold symmetry axes兲, thus saving (AlCp* ) 4 c D 2d 1 1
E BP SV共P兲 796 2:33 2:30
a factor of approximately the order of the point group in a
Open shell.
computation time and disk space requirements. Spin symme- b
4-共dimethylamino兲benzonitrile.
try is also taken into account for spin restricted closed shell c
Cp * ⫽pentamethylcyclopentadienyl.

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002 TDDFT for excited state properties 7439

TABLE III. Basis set dependence of the calculated properties of the 1 1 A ⬙ TABLE IV. Basis sets of aug-TZVPP quality for the elements Li, Be, Mg,
excited state of CH2 O. The TDHF method is used. Energies E are in Har- Sc, and Cu. We list the names of the corresponding standard basis sets from
tree, excitation energies T e in eV, bond lengths in pm, out-of-plane angles ␾ the TURBOMOLE library 共see Refs. 42, 43, and 62兲 as well as the primitive
in degrees, and dipole moments ␮ in D. Gaussians added. Exponents obtained from downward extrapolation are des-
ignated by E.
Basis E Te C–O C–H ␾ ␮
Augmentation
SVP ⫺113.620 440 4.336 125.33 109.43 26.14 1.411
aug-SVP ⫺113.629 459 4.356 125.04 109.38 24.21 1.350 Element Basis l Exponent Remarks
aug-TZVP ⫺113.754 481 4.328 125.25 108.68 25.56 1.290
Li TZVPP s 0.010 872 E
TZVPP ⫺113.758 247 4.365 124.96 108.59 25.89 1.324
p 0.009 000 E
aug-TZVPP ⫺113.759 062 4.363 124.86 108.61 24.45 1.285
d 0.003 000
aug-cc-pVQZ ⫺113.762 254 4.372 124.75 108.57 24.47 1.282
d 0.001 000

Be TZ.3P s 0.023 096 E

p 0.016 372 E
d 0.510 000 Ref. 63
⭐10⫺6 Hartree, maximum gradient norm ⭐10⫺3 a.u.). The d 0.127 500 Ref. 63
ground state energy and density matrix were converged to at Mg TZVDP s 0.013 457 E
least 10⫺7 a.u.; fine quadrature grids 关size 4 共Ref. 36兲兴 were p 0.015 285 E
used. Harmonic vibrational frequencies and zero point ener- d 0.350 000 Ref. 63
gies 共ZPEs兲 were determined by numerical differentiation of d 0.087 500 Ref. 63
the analytic gradients 共central differences, default displace- Sc TZVDP s 0.011 775 E
ments 0.02 bohr兲. T e values were obtained as energy differ- p 0.010 730 E
ences between the excited and ground states in their equilib- d 0.040 981 E
rium structures, T 0 values include ZPE differences. All f 0.350 000
f 0.090 000
calculated T 0 values, structural parameters, dipole moments,
g 0.100 000
and frequencies were obtained in the harmonic approxima-
tion. Cu TZVP s 0.013 896 E
In order to assess the basis set dependence of the calcu- p 0.032 555 E
d 0.084 703 E
lated excited state properties, we have optimized the 1 1 A ⬙ f 0.600 000
state of CH2 O in various basis sets, see Table III. SV denotes f 0.200 000
split valence42 and TZV triple zeta valence43 basis sets; P is a g 0.400 000
single set of polarization functions, and PP a double one,
e.g., 2d1 f for elements C–F. The prefix ‘‘aug’’ stands for
sets of diffuse basis functions44 containing an additional
primitive Gaussian for each l-quantum number in the origi-
nal basis set. The aug-cc-pVQZ basis set45,46 is of quadruple dard TURBOMOLE basis sets with diffuse and polarization
zeta valence quality, with three sets of polarization functions functions. The resulting basis sets are listed in Table IV.
and diffuse augmentation. The results in Table III were ob-
tained using TDHF, but the other methods are expected to
exhibit very similar basis set dependence.
The diffuse augmentation has little influence on the en- V. RESULTS
ergy of the rather compact 1 1 A ⬙ state, but is necessary for an A. Choice of the test set
accurate description of the geometry. Considering the con-
Compared to ground states, the amount of accurate ex-
vergence of both energy and properties, a hierarchy of basis
perimental data on excited states in the literature is very lim-
sets for excited state calculations is
ited, especially for polyatomic molecules. The choice of mol-
aug-SVP⬍aug-TZVPP⬍aug-cc-pVQZ. 共34兲 ecules for the test set was therefore mainly dictated by the
availability of 共presumably兲 accurate results from gas-phase
This is not unexpected from the experience with ground state experiments. Moreover, the test set was designed to contain a
calculations and other response properties such as polariz- variety of excited states of different character and spin mul-
abilities. The energy convergence indicates that the aug-cc- tiplicity. The set thus should be considered a collection of
pVQZ results are close to the basis set limit. From Table III, interesting examples. Although tempting, a statistical analy-
we estimate the basis set errors for aug-TZVPP to be ap- sis does not make sense, because errors are of very different
proximately ⫾0.02 eV in the excitation energy, ⫾0.2 pm in origin and show systematic behavior. We have observed that,
bond lengths and ⫾0.02 D in the dipole moment 共for the by adding or removing a few cases, the outcome of a statis-
1 1 A ⬙ state of CH2 O). Since this appears quite sufficient for tics can be changed rather arbitrarily. Moreover, a statistical
our present purpose, all calculations were performed using analysis does not reflect the different quality of errors; for
the aug-TZVPP basis set, unless stated otherwise. For the example, a systematic overestimation of element–hydrogen
elements Li, Be, Mg, Sc, and Cu, Dunning diffuse functions bond lengths is much less severe in practice than a failure to
are not available. We, therefore, have constructed basis sets describe the strong C–O bond elongation in carbonyls upon
of quality comparable to aug-TZVPP by augmenting stan- n- ␲ * excitation. For these reasons, we do not give any sta-

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 7440 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002
tistical data in the following, except for maximum deviations
from experiment, indicated by bold typeface in Tables
V–VIII.
We do not want to conceal that for certain manifest mul-
tireference cases, such as the 1 2 B 2 state of NO2 , all meth-
ods studied here exhibit instabilities of the reference
state.47,48 This means that the ground state intersects with an
excited state of different symmetry, usually during geometry
optimization. As a consequence, negative or imaginary exci-
tation energies occur, and the response formalism breaks
down.22 Another well-known example is the singlet ground
state of the C2 molecule, which is triplet unstable in HF and
DFT methods even at the ground state geometry. An ad-
equate treatment of these difficult systems remains the do-
main of multideterminantal methods.
B. Adiabatic excitation energies
Adiabatic excitation energies are in general more diffi-
cult to calculate than vertical ones because they require a
theoretical model to work well not only at the equilibrium
structure of the ground state but also of the excited state.
Obviously, the CIS and TDHF results in Table V depend
sensitively on the quality of the HF reference determinant in
the excited state structure. In cases such as the 1 1 ⌸ states of
CO and SiO, the 2 1 B 1 state of H2 O, the 1 1 A ⬙ state of
CH2 O, or the 1 1 B 2u state of C6 H6 , the structure changes
little upon excitation, and the HF reference is reasonable for
both ground and excited states. Thus, as expected from varia-
tional considerations, CIS adiabatic excitation energies are
1–2 eV too high, which is somewhat improved by TDHF. On
the other hand, for the 1 1 ⌺ ⫺ state of N2 , the 1 1 A ⬙ state of
HCN, or the 1 1 A u state of C2 H2 , the CIS adiabatic excita-
tion energies are too small by about 1 eV, indicating multi-
reference ground states at the excited state structure. TDHF
fails even more badly due to nonreal instabilities leading to
imaginary excitation energies for excited singlet states of N2
and C2 H2 . Nonreal instabilities of the HF reference are not
uncommon for geometries differing significantly from that of
the ground state, while for triplet states TDHF stability prob-
lems often occur for vertical excitations already. This lack of
stability makes TDHF rather useless as a method for the
investigation of excited potential energy surfaces.
Absolute errors in the LDA and GGA adiabatic excita-
tion energies are considerably smaller and more systematic
than those found for CIS and TDHF. For example, LDA as
well as the BP86 and PBE GGAs give a correct ordering of
the three lowest singlet excited states of N2 despite the dif-
ferent structure and character of these states. Even for the
difficult cases HCN and C2 H2 , comparably small errors of
⫺0.4 or ⫺0.5 eV are observed. This can be attributed to the
fact that the ground state KS solution is stable on a much
larger region of the potential energy surface than the HF
solution.48 The errors in the LDA and GGA adiabatic excita-
tion energies follow a very similar pattern as those observed
in vertical excitation energies for these methods. There is a
tendency to underestimate excitation energies systematically,
which increases as the KS ionization threshold is reached.
Excitation energies of Rydberg states such as 2 1 B 1 in H2 O
are calculated too small, and may become even smaller if the
Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
F. Furche and R. Ahlrichs
basis set is further augmented by diffuse functions. In the
LDA and GGA treatments, 2 1 B 1 of H2 O is a continuum
state, which is clearly unphysical; these problems are well
known from vertical excitation energy calculations, and vari-
ous remedies have been suggested.49,50 Hybrid functionals
perform significantly better for diffuse states, which is also
not unexpected.51 As for vertical excitations, gradient correc-
tions do not improve upon the LDA adiabatic excitation en-
ergies in general. The hybrid results mostly interpolate be-
tween TDHF and the GGA excitation energies. Error
compensation therefore is observed only if GGA excitation
energies are somewhat too small, and the TDHF values are
too large. In all other situations, the admixture of HF ex-
change does not improve the GGA results. This is especially
true for states where TDHF exhibits instabilities. The TDA
scheme11 can be expected to be more robust, in particular for
triplet states, whose excitation energy is often severely un-
derestimated by TDDFT methods; its performance in predict-
ing adiabatic excitation energies may deserve further atten-
tion. Considering the present results, we hesitate to
recommend hybrid functionals as the first choice in excited
state calculations, since it appears to be difficult to decide a
priori whether error compensation works for a particular
state or not.
C. Structures, dipole moments, and vibrational
frequencies
Relative errors in computed excited state structures, di-
pole moments, and vibrational frequencies are for all meth-
ods much smaller than errors in the excitation energies, as
can be seen from Tables VI–VIII. In fact, the accuracy is
often comparable to that of the corresponding ground state
methods, as long as no stability problems occur. CIS bond
lengths are in general several pm too small, and vibrational
frequencies are too high by 20% and more, error character-
istics well known from HF ground state calculations. One
might conclude that CIS is already a useful method for the
investigation of excited states. There are drawbacks, how-
ever. CIS apparently tends to an artificial symmetry breaking
in delocalized systems such as the 1 1 B 2 state of pyridine,5 as
is also familiar from ground state HF. The CIS geometry is
qualitatively in error for the 1 3 A ⬙ state of CH2 S, which is
predicted planar. The C–O stretching mode ␯ 2 in the n- ␲ *
excited carbonyl compounds CH2 O and (CHO) 2 is
400– 600 cm⫺1 too high in the CIS treatment, and C–O bond
lengths are too small. TDHF is not reliable due to stability
problems and hardly improves upon CIS even for stable sys-
tems.
All density functional methods yield substantially better
results for excited state structures, dipole moments, and vi-
brational frequencies than CIS. As for ground states, bond
lengths are somewhat overestimated by GGA functionals; the
LDA and hybrid functionals give slightly shorter bonds.
Nevertheless, the GGA errors are systematic, and the general
picture obtained from GGA calculations is very consistent.
For example, the elongation of the C–O bond upon excita-
tion in CH2 O and (CHO) 2 is best described by the GGA, and
the same trend is observed for other excitations with consid-
erable changes in the bond order. The accuracy of the calcu-
 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002 TDDFT for excited state properties 7441

TABLE V. Calculated adiabatic excitation energies 共eV兲 compared to experiment. Calculated and experimental
energies include ZPE differences (T 0 values兲 except for diatomics and porphyrin (T e values兲. Experimental
results are from Ref. 64 for diatomic and from Ref. 65 for polyatomic molecules, unless otherwise stated.
Instabilities of the reference state are denoted by 共I兲; bold typeface indicates maximum absolute deviations from
experiment. RPA stands for TDHF, LDA for PW, BLP for BLYP, BP for BP86, PB0 for PBE0 and B3P for
B3LYP.

System State CIS RPA LDA BLP BP PBE PB0 B3P Expt.
BeH 1 2⌸ 2.72 2.69 2.35 2.53 2.51 2.49 2.52 2.55 2.48

Li2 1 1⌺ ⫹
u 2.12 1.96 1.91 1.92 1.88 1.91 1.95 1.95 1.74

BH 1 ⌸ 1
2.85 2.64 2.49 2.71 2.72 2.66 2.64 2.68 2.87

NH 1 ⌸ 3
4.18 4.02 3.63 3.91 4.02 3.98 3.96 3.87 3.70

BeO 1 1⌸ 2.37 2.27 1.26 1.05 1.21 1.15 1.44 1.26 1.17

BF 1 1⌸ 6.55 6.22 5.95 5.99 6.04 6.02 6.12 6.08 6.34

N2 1 3
⌸g 7.70 7.20 7.03 6.76 6.73 6.74 6.96 6.98 7.39
1 1
⌺⫺
u 7.54 共I兲 8.27 7.99 8.09 8.10 7.95 7.90 8.45
1 1
⌸g 9.60 9.27 8.46 8.34 8.39 8.39 8.69 8.61 8.59
1 1
⌬u 8.12 7.59 8.86 8.32 8.50 8.55 8.55 8.36 8.94

CO 1 3⌸ 5.77 5.13 5.73 5.49 5.40 5.43 5.49 5.59 6.04

1 1⌸ 8.81 8.50 7.84 7.79 7.83 7.82 8.06 8.00 8.07

NO 1 2⌺ ⫹ 共I兲 共I兲 共I兲 5.50 5.75 5.45 6.22 6.09 5.45

Mg2 1 1⌺ ⫹
u 3.33 3.23 3.20 3.21 3.20 3.16 3.17 3.24 3.23

SiO 1 1⌸ 6.11 5.87 5.06 4.93 4.98 4.97 5.25 5.17 5.31

ScO 1 ⌸ 2
2.00 1.94 2.00 1.94 1.93 1.86 1.87 1.95 2.04a

CuH 2 1⌺ ⫹ 3.93 3.84 3.06 2.85 2.97 2.91 3.09 2.99 2.91

H2 O 2 1B 1 11.73 11.70 9.01 8.70 8.84 8.88 9.82 9.50 10.00

2
PH2 1 A1 2.78 2.59 2.13 2.34 2.40 2.38 2.40 2.34 2.27

SO2 1 B13
3.01 共I兲 2.73 2.43 2.43 2.45 2.51 2.54 3.19

SiF2 1 1B 1 5.91 5.58 5.15 5.11 5.11 5.09 5.29 5.29 5.34

CCl2 1 1B 1 2.17 共I兲 共I兲 2.04 2.00 1.96 共I兲 1.94 2.14b

CS2 1 3A 2 3.37 共I兲 3.23 2.98 3.02 3.05 3.04 3.01 3.25

HCN 1 1A ⬙ 5.64 4.88 6.27 5.97 6.07 6.09 6.03 5.96 6.48

HCP 1 1A ⬙ 3.55 共I兲 4.08 3.86 3.95 3.96 3.88 3.82 4.31

C2 H2 1 Au1
4.52 共I兲 4.98 4.69 4.77 4.79 4.77 4.69 5.23

CH2 O c
1 1A ⬙ 4.44 4.25 3.38 3.41 3.44 3.42 3.64 3.60 3.49
1 3A ⬙ 3.53 共I兲 2.68 2.57 2.54 2.53 2.70 2.73 3.12

CH2 Sc 1 1A 2 2.61 2.50 1.94 1.94 1.98 1.97 2.11 2.07 2.03
1 3A ⬙ 1.88 共I兲 1.44 1.36 1.34 1.35 1.43 1.45 1.80

trans-(CHO) 2 1 1A u 3.59 3.41 1.90 2.04 2.01 1.98 2.45 2.42 2.72

Benzened 1 1 B 2u 5.87 5.67 5.01 4.84 4.91 4.91 5.24 5.13 4.72e

Pyridinef 1 1B 2 5.89g 5.57g 5.11g 4.94 5.00 5.01 5.36 5.24 4.76h

Porphyrin 1 1 B 1u 2.45 1.76 2.19 2.08 2.11 2.12 2.31 2.25 2.02i
a
Reference 66.
b
Reference 67.
c
Experimental data from Ref. 68.
d
ZPE calculated in the aug-SVP basis.
e
Reference 69.
f
Structures and ZPE calculated in the SVP basis.
g
Saddle point.
h
Reference 70.
i
T e value obtained by correction of experimental T 0 value 共Ref. 71兲 by ZPE difference calculated using B3LYP/
SVP.

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 7442 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002 F. Furche and R. Ahlrichs

TABLE VI. Calculated excited state equilibrium structures compared to experiment 共bond lengths in pm, angles
in degrees兲. Experimental r e values are corrected for anharmonicity. See Table V for further explanations.

System State Param. CIS RPA LDA BLP BP PBE PB0 B3P Expt.

BeH 1 ⌸ 2
re 132 132 134 133 133 134 133 132 133
1 ⫹
Li2 1 ⌺u re 309 320 307 307 307 307 307 307 311

BH 1 1⌸ re 120 120 123 121 122 123 121 121 122

NH 1 ⌸ 3
re 102 101 106 106 105 105 104 104 104

BeO 1 ⌸ 1
re 142 144 143 143 145 144 143 143 146

BF 1 1⌸ re 129 129 129 130 130 130 130 130 130

N2 1 3
⌸g re 118 119 120 121 121 121 120 120 121
1 1
⌺⫺
u re 126 共I兲 127 129 129 129 128 128 128
1 1
⌸g re 119 120 121 122 122 122 121 121 122
1 1
⌬u re 123 126 127 129 129 128 126 127 127

CO 1 3⌸ re 118 119 120 121 121 121 120 120 121

1 1⌸ re 121 122 122 124 124 123 122 123 124

NO 1 2⌺ ⫹ re 共I兲 共I兲 共I兲 107 107 107 105 105 106

1 ⫹
Mg2 1 ⌺u re 324 321 309 317 319 316 316 316 308

SiO 1 1⌸ re 158 159 160 163 162 162 161 162 162

ScO 1 ⌸ 2
re 166 166 168 172 171 170 168 170 169a

CuH 2 1⌺ ⫹ re 151 150 159 163 162 162 157 158 157

H2 O 2 1B 1 O–H 96 96 102 102 102 101 100 100 102

⬔(HOH) 109 109 105 104 105 104 105 105 107

PH2 1 2A 1 P–H 139 139 141 141 141 141 140 140 140
⬔(HPH) 120 121 123 122 122 122 122 122 123

SO2 1 3B 1 S–O 144 共I兲 151 154 153 153 150 151 149
⬔(OSO) 127 共I兲 126 125 126 126 127 126 126

SiF2 b 1 1B 1 Si–F 158 158 163 167 166 166 162 163 160
⬔(FSiF) 111 112 114 114 114 114 114 114 116

CCl2 b 1 1B 1 C–Cl 165 共I兲 共I兲 167 166 165 共I兲 165 165
⬔(ClCCl) 133 共I兲 共I兲 134 135 135 共I兲 135 131

CS2 1 3A 2 C–S 159 共I兲 163 165 164 164 162 163 164
⬔(SCS) 145 共I兲 138 137 137 137 138 138 136

HCN 1 1A ⬙ C–H 110 108 115 114 114 114 112 112 114
N–C 128 135 129 130 130 130 129 130 130
⬔(HCN) 121 125 123 123 123 123 123 123 125

HCP 1 1A ⬙ P–C 168 共I兲 169 172 171 171 170 170 169
⬔(HCP) 132 共I兲 129 127 127 128 130 129 128

C2 H2 1 1A u C–C 135 共I兲 136 137 137 137 136 137 139
⬔(HCC) 125 共I兲 124 122 122 122 122 123 120

CH2 Oc 1 1A ⬙ C–H 109 109 111 110 110 110 110 109 110
C–O 125 125 128 131 131 131 128 129 132
⬔(HCH) 118 117 115 117 116 116 116 117 118
␾ 22 24 34 31 33 33 31 30 34
1 3A ⬙ C–H 109 共I兲 112 111 110 111 111 110 108
C–O 124 共I兲 128 132 131 131 128 129 131
⬔(HCH) 112 共I兲 111 112 110 111 110 112 122
␾ 40 共I兲 44 46 49 48 47 45 41

CH2 Sc 1 1A 2 C–H 107 107 109 109 109 109 108 108 108
C–S 163 163 168 171 170 170 167 168 168
⬔(HCH) 119 118 120 120 120 120 119 119 121
1 3A ⬙ C–H 107 共I兲 109 109 109 109 108 108 108
C–S 162 共I兲 167 171 170 170 167 168 168
⬔(HCH) 119 共I兲 120 119 118 118 119 119 119

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002 TDDFT for excited state properties 7443

TABLE VI. 共Continued.兲

System State Param. CIS RPA LDA BLP BP PBE PB0 B3P Expt.

␾ 0 共I兲 1 22 26 24 18 15 12
d 1
trans-(CHO) 2 1 Au C–H 109 109 112 111 111 111 110 110 112
C–O 119 120 123 125 124 124 122 123 125
C–C 151 151 146 148 148 147 148 148 146
⬔(HCC) 112 113 113 114 114 114 113 113 114
⬔(OCC) 124 123 126 125 126 126 125 125 124

Benzenee 1 1 B 2u C–H 107 107 109 109 109 109 108 108 108
C–C 141 141 142 143 143 143 142 142 143
a
Reference 66.
b
Experimental data from Ref. 67.
c
Experimental data from Ref. 68.
d
Experimental data from Ref. 72.
e
Experimental data from Ref. 69.

lated excited state dipole moments is surprisingly high for
both LDA and GGA, while hybrid functionals can suffer
from artifacts of TDHF and do not always improve upon
GGA. For vibrational frequencies of excited states, GGA
functionals, especially BP86 and PBE, are superior to all
other methods. Some caution is appropriate, however, since
most experimental frequencies in Table VIII are not cor-
rected for anharmonicity effects. Again, there is apparently
no benefit from admixing HF exchange. The LDA frequen-
cies tend to somewhat lower values. For example, the lowest
b 1 mode of the 1 1 B 2 state of pyridine is imaginary in the
LDA, which is probably an artifact. On the other hand, the
critical C–O stretching modes in the lowest singlet excited
states of CH2 O and (CHO) 2 are calculated about 100 cm⫺1
higher than within the GGA. However, even the GGA meth-
ods fail to give the correct ordering of the totally symmetric
␯ 2 and ␯ 3 modes. With a measured separation of only 100
and 200 cm⫺1 , the correct assignment of ␯ 2 and ␯ 3 is cer-
tainly a challenge. This example illustrates well the degree of
accuracy that can be expected from TDDFT excited state
calculations. In the vast majority of the vibrations in Table
TABLE VII. Calculated excited state dipole moments (D) compared to
experiment. See Table V for further explanations.
System State CIS RPA LDA BLP BP
BH 1 ⌸1
0.63 0.48 0.51 0.44 0.52 0.49 0.49 0.49 0.58
VIII, the GGA assignment is correct, though. This is most
important for the possible use of TDDFT in excited state
structure elucidation, since vibrational frequencies are often
the only structural information which is experimentally ac-
cessible in excited states of larger systems.
VI. CONCLUSIONS
Stability of the ground state is critical in single-reference
treatments of excited state properties. TDDFT is consider-
ably more robust than CIS and especially TDHF since the
KS reference is stable on a larger range of nuclear geom-
etries than the HF reference. TDDFT adiabatic excitation en-
ergies improve upon CIS and TDHF, but errors can reach 0.4
eV and more, which is still far from the desired ‘‘chemical
accuracy’’ of 0.05 eV. However, TDDFT excited state struc-
tures, dipole moments, and force constants are nearly as ac-
curate as the corresponding DFT ground state properties;
similar behavior can be expected for activation barriers and
reaction paths. In view of the success of ground state DFT,
this is a promising result. With the availability of stable and
efficient TDDFT implementations, CIS will become obsolete
for excited state geometry optimizations since it generally
leads to inferior results at the same computational cost. We
recommend TDDFT for standard applications to excited
PBE PB0 B3P Expt. states of molecules in the size range of 20–200 atoms, where
correlated ab initio methods are still too expensive.

NH 1 ⌸3
1.26 1.22 1.37 1.31 1.29 1.28 1.27 1.30 1.31
ACKNOWLEDGMENTS
CO 1 ⌸3
2.14 2.52 1.24 1.36 1.34 1.30 1.58 1.57 1.37
1 1⌸ 1.08 1.08 0.29 0.26 0.25 0.23 0.50 0.49 0.34a This work was supported by the Deutsche Forschungs-
NO 1 ⌺2 ⫹
共I兲 共I兲 共I兲 1.21 1.71 1.54 1.48 1.17 1.10 gemeinschaft, Sonderforschungsbereich 195 共‘‘Lokalisierung
von Elektronen in makroskopischen und mikroskopischen
ScO 1 ⌸2
4.96 5.09 4.36 4.69 4.34 4.33 4.54 4.88 4.14b
Systemen’’兲, and by the Fonds der Chemischen Industrie.
CH2 Oc 1 1 A ⬙ 1.37 1.28 1.63 1.56 1.56 1.54 1.51 1.53 1.57
1 3A ⬙ 1.40 共I兲 1.57 1.41 1.42 1.42 1.42 1.41 1.29
APPENDIX A: DERIVATION OF THE EQUATIONS
CH2 Sc 1 1 A 2 0.44 0.37 0.84 0.80 0.84 0.83 0.73 0.72 0.79
FOR THE LAGRANGE MULTIPLIERS Z AND W
1 3A ⬙ 0.02 共I兲 0.56 0.50 0.56 0.55 0.41 0.38 0.57
a In this Appendix, it is shown in detail how to derive the
Reference 73.
b
Reference 66. equations for the Lagrange multipliers Z and W in Sec. II C
c
Experimental data from Ref. 68. from the orbital stationarity condition 共15兲. In order to trans-

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 7444 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002 F. Furche and R. Ahlrichs

TABLE VIII. Calculated harmonic vibrational frequencies (cm⫺1 ) compared to experiment. Experimental ␻ e
values are corrected for anharmonicity. See Table V for further explanations.

System State Mode CIS RPA LDA BLP BP PBE PB0 B3P Expt.

BeH 1 ⌸ 2
␻e 2241 2232 2083 2117 2096 2090 2150 2161 2089
1 ⫹ ␻e
Li2 1 ⌺u 271 235 249 243 256 253 258 248 255

BH 1 1⌸ ␻e 2551 2602 2281 2306 2263 2267 2389 2406 2251

NH 1 ⌸ 3
␻e 3594 3618 3015 2937 3000 3015 3219 3123 3231

BeO 1 ⌸ 1
␻e 1254 1084 1250 1218 1210 1205 1219 1251 1144

BF 1 1⌸ ␻e 1369 1368 1308 1227 1244 1244 1297 1274 1265

N2 1 3
⌸g ␻e 2027 1781 1858 1728 1767 1782 1840 1792 1733
1 1
⌺⫺
u ␻e 1686 共I兲 1387 1282 1311 1318 1545 1513 1530
1 1
⌸g ␻e 1904 1802 1789 1665 1707 1719 1789 1736 1694
1 1
⌬u ␻e 1861 1496 1544 1457 1483 1488 1622 1576 1559

CO 1 3⌸ ␻e 1963 1854 1831 1729 1762 1769 1836 1792 1743

1 1⌸ ␻e 1636 1584 1611 1474 1513 1524 1591 1543 1518

NO 1 2⌺ ⫹ ␻e 共I兲 共I兲 共I兲 2393 2397 2443 2574 2552 2374

Mg2 1 1⌺ ⫹
u ␻e 153 160 177 153 156 153 163 156 191

SiO 1 1⌸ ␻e 944 907 1057 874 895 897 925 900 853

ScO 1 ⌸ 2
␻e 973 967 865 809 837 845 891 846 876a

CuH 2 1⌺ ⫹ ␻e 1817 1811 1584 1588 1604 1611 1738 1711 1698

H2 O 2 1B 1 ␯ 1 (a 1 ) 3832 3834 2995 2978 2946 2997 3261 3213 3170

PH2 1 2A 1 ␯ 2 (a 1 ) 1046 1042 919 951 934 929 959 977 951

SO2 1 B13
␯ 1 (a 1 ) 1157 共I兲 935 763 881 888 973 928 906

␯ 2 (a 1 ) 454 共I兲 353 338 341 342 368 361 360

SiF2 1 1B 1 ␯ 1 (a 1 ) 866 870 702 623 648 646 744 715 598
␯ 2 (a 1 ) 300 296 230 220 223 221 249 245 342

CCl2 b 1 1B 1 ␯ 1 (a 1 ) 678 共I兲 共I兲 590 617 622 共I兲 630 634
␯ 2 (a 1 ) 311 共I兲 共I兲 280 273 260 共I兲 272 303

CS2 1 3A 2 ␯ 1 (a 1 ) 716 共I兲 716 665 681 686 705 689 692
␯ 2 (a 1 ) 279 共I兲 309 305 306 305 296 303 311

HCN 1 1A ⬙ ␯ 2 (a ⬘ ) 629 425 950 973 969 964 983 992 941
␯ 3 (a ⬘ ) 1766 1686 1532 1458 1477 1483 1557 1531 1496

HCP 1 1A ⬙ ␯ 2 (a ⬘ ) 588 共I兲 668 708 699 690 668 694 567
␯ 3 (a ⬘ ) 844 共I兲 966 914 931 936 983 959 951

C2 H2 1 1A u ␯ 2 (a g ) 1549 共I兲 1448 1389 1427 1403 1456 1437 1385

␯ 3 (a g ) 1149 共I兲 1041 1084 1098 1078 1088 1097 1048

CH2 Oc 1 1A ⬙ ␯ 1 (a ⬘ ) 3135 3134 2871 2918 2906 2916 3002 2987 2846
␯ 2 (a ⬘ ) 1633 1620 1364 1279 1289 1294 1408 1361 1183
␯ 3 (a ⬘ ) 1385 1390 1194 1247 1249 1254 1287 1301 1293
␯ 5 (a ⬙ ) 3223 3217 2963 3023 3008 3016 3096 3083 2968
␯ 6 (a ⬙ ) 947 954 809 870 852 855 873 890 904
1 3A ⬙ ␯ 2 (a ⬘ ) 1613 共I兲 1369 1201 1251 1261 1369 1323 1283

CH2 Sc 1 1A 2 ␯ 1 (a 1 ) 3265 3265 3050 3063 3066 3068 3148 3133 3034
␯ 2 (a 1 ) 1488 1485 1293 1334 1329 1311 1365 1377 1316
␯ 3 (a 1 ) 1070 1071 904 835 862 870 942 903 820
␯ 5 (b 2 ) 3366 3364 3162 3176 3180 3183 3260 3242 3081
␯ 6 (b 2 ) 851 846 749 773 773 752 789 803 799
1 3A ⬙ ␯ 2 (a ⬘ ) 1482 共I兲 1286 1329 1306 1303 1346 1364 1320
␯ 3 (a ⬘ ) 1096 共I兲 912 836 858 868 948 909 859

trans-(CHO) 2 d 1 1A u ␯ 1 (a g ) 3164 3177 2837 2888 2882 2883 2986 2965 2809
␯ 2 (a g ) 1806 1796 1589 1475 1513 1520 1622 1570 1391
␯ 3 (a g ) 1295 1305 1170 1222 1208 1204 1226 1237 1195

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002 TDDFT for excited state properties 7445

TABLE VIII. 共Continued.兲

System State Mode CIS RPA LDA BLP BP PBE PB0 B3P Expt.

␯ 4 (a g ) 1009 1017 987 957 970 973 991 972 952

␯ 5 (a g ) 544 544 512 509 508 509 521 520 509
␯ 6 (a u ) 838 838 757 761 763 759 796 786 720
␯ 7 (a u ) 256 256 258 249 253 252 267 259 233
␯ 8 (b g ) 859 856 767 765 765 764 797 789 735
␯ 10(b u ) 1709 1682 1565 1460 1492 1499 1576 1530 1281
␯ 11(b u ) 1326 1331 1093 1125 1112 1110 1197 1198 1172
␯ 12(b u ) 441 441 337 344 341 341 397 392 379

Benzenee 1 1 B 2u ␯ 1 (a 1g ) 3392 3392 3146 3147 3155 3161 3252 3226 3093
␯ 2 (a 1g ) 1026 1025 965 928 939 943 981 963 923
␯ 3 (a 2g ) 1455 1452 1231 1286 1286 1275 1332 1331 1327
␯ 4 (a 2u ) 647 647 575 592 575 573 596 619 515-518
␯ 7 (b 2g ) 905 907 868 868 842 838 879 922 745
␯ 8 (b 2g ) 426 434 247 370 331 319 377 418 365
␯ 9 (b 2u ) 1849 1912 1458 1380 1408 1415 1532 1482 1571
␯ 10(b 2u ) 1254 1254 1099 1126 1125 1119 1161 1160 1150
␯ 11(e 1g ) 724 724 595 629 606 602 639 666 581
␯ 12(e 1u ) 3381 3381 3137 3137 3145 3152 3242 3217 3084
␯ 13(e 1u ) 1570 1570 1374 1386 1388 1383 1449 1440 1405
␯ 14(e 1u ) 1062 1061 944 920 933 934 989 971 920
␯ 15(e 2g ) 3367 3366 3125 3125 3132 3139 3229 3204 3077
␯ 16(e 2g ) 1713 1714 1545 1501 1520 1523 1604 1570 1516
␯ 17(e 2g ) 1249 1248 1112 1128 1132 1128 1169 1163 1148
␯ 18(e 2g ) 571 560 495 512 507 506 523 525 521
␯ 19(e 2u ) 887 888 798 797 781 779 820 843 717
␯ 20(e 2u ) 269 286 262 275 272 267 292 304 238

Pyridinef 1 1B 2 ␯ 6 (a 1 ) 1309 1311 1172 1169 1173 1173 1221 1210 1215g
␯ 8 (a 1 ) 1086 1089 971 939 950 954 1003 984 995
␯ 9 (a 1 ) 1049 1048 941 911 921 925 977 960 950g
␯ 10(a 1 ) 565 566 489 502 496 495 512 516 542
␯ 13(a 2 ) 共im.兲 共im.兲 257 272 266 264 313 309 327g
a
Reference 66.
b
Experimental data from Ref. 67.
c
Experimental data from Ref. 68.
d
Experimental data from Ref. 74.
e
Aug-SVP basis set, experimental data from Ref. 69.
f
SVP basis.
g
Reference 70.

form Eq. 共15兲 to a more manageable form, we multiply both Q i j ␴ ⫹H ⫹

i j ␴ 关 Z 兴 ⫽W i j ␴ , 共A4兲
sides by C ␮ q ␴ and sum over ␮. Using the definition of L
共12兲, this leads to where H is defined in Eq. 共20兲, and, using the unrelaxed
difference density matrix T from Eq. 共19兲,
⳵ F ia ␴ ⬘
Q pq ␴ ⫹ 兺 Z ia ␴ ⬘ 兺␮ ⳵ C ␮p␴ ␮q␴
C
ia ␴ ⬘ Q i j␴⫽ 兺a ⍀ 兵 共 X⫹Y 兲 ia ␴共 X⫺Y 兲 ja ␴共 X⫺Y 兲 ia ␴共 X⫹Y 兲 ja ␴ 其
⳵S
⫽ 兺 W rs ␴ ⬘ 兺
rs ␴ ⬘
C ␮q␴ , 共A1兲
⬘
rs ␴ ,r⭐s ␮ ⳵ C ␮p␴ ⫺ 兺a ⑀ a ␴ 兵 共 X⫹Y 兲 ia ␴共 X⫹Y 兲 ja ␴
where Q is given by
⫹ 共 X⫺Y 兲 ia ␴ 共 X⫺Y 兲 ja ␴ 其 ⫹H ⫹
i j␴关 T 兴
⳵ G 关 X,Y ,⍀ 兴
Q pq ␴ ⫽ 兺␮ ⳵ C ␮p␴
C ␮q␴ . 共A2兲
⫹2 兺 xc
g i j ␴ kc ␴ ⬘ ld ␴ ⬙ 共 X⫹Y 兲 kc ␴ ⬘ 共 X⫹Y 兲 ld ␴ ⬙ . 共A5兲
In the next step, we evaluate the derivatives with respect to ⬘
kc ␴ ld ␴ ⬙
C. This is conveniently done by expressing all operator ma- Here, the definition of G 共1兲 has been inserted; moreover, the
trix elements in terms of AO integrals, e.g., fact that the MO coefficients C satisfy the KS equations and
the excitation vector solves the TDKS eigenvalue problem
h pq ␴ ⫽ 兺
␮␯
C ␮p␴h ␮␯C ␯q␴ . 共A3兲 共6a兲 has been exploited, which is clearly legitimate after the
derivative has been taken. Similarly, we obtain for the occ–
If p⫽i⭐q⫽ j are both occupied, Eq. 共A1兲 thus transforms to virt block

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 7446 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002 F. Furche and R. Ahlrichs

⫹
Q ia ␴ ⫹ ⑀ a ␴ Z ia ␴ ⫹H ia ␴ 关 Z 兴 ⫽W ia ␴ , 共A6兲 R. Burcl, R. D. Amos, and N. C. Handy, Chem. Phys. Lett. 355, 8 共2002兲.
15
16
L. N. Oliveira, E. K. U. Gross, and W. Kohn, Phys. Rev. A 37, 2821
共1988兲.
Q ia ␴ ⫽ 兺b 兵 共 X⫹Y 兲 ib ␴ H ab⫹ ␴ 关 X⫹Y 兴 17
18
A. Görling, Phys. Rev. A 59, 3359 共1999兲.
M. Levy and A. Nagy, Phys. Rev. Lett. 83, 4361 共1999兲.
19
⫺ ⫹
I. Frank, J. Hutter, D. Marx, and M. Parrinello, J. Chem. Phys. 108, 4060
⫹ 共 X⫺Y 兲 ib ␴ H ab ␴ 关 X⫺Y 兴 其 ⫹H ia ␴ 关 T 兴 共1998兲.
20
E. R. Davidson and L. E. Nitzsche, J. Am. Chem. Soc. 101, 6524 共1979兲.
21
J. V. Ortiz, J. Chem. Phys. 101, 6743 共1994兲.
⫹2 兺 xc
g ia ␴ jb ␴ ⬘ kc ␴ ⬙ 共 X⫹Y 兲 jb ␴ ⬘ 共 X⫹Y 兲 kc ␴ ⬙ , 22
F. Furche, J. Chem. Phys. 114, 5982 共2001兲.
⬘
jb ␴ kc ␴ ⬙ 23
A. D. Becke, J. Chem. Phys. 98, 1372 共1993兲.
共A7兲
24
A. D. Becke, J. Chem. Phys. 98, 5648 共1993兲.
25
A. D. McLachlan and M. A. Ball, Rev. Mod. Phys. 36, 844 共1964兲.
26
for the virt–occ block Geometrical Derivatives of Energy Surfaces and Molecular Properties,
Vol. 166 of NATO ASI series. Series C, Mathematical and Physical Sci-
Q ai ␴ ⫹ ⑀ i ␴ Z ia ␴ ⫽W ia ␴ , 共A8兲 ences, edited by P. Jørgensen and J. Simons 共Reidel, Dordrecht, 1985兲.
27
R. M. Sternheimer and H. M. Foley, Phys. Rev. 92, 1460 共1953兲.

兺j 兵 共 X⫹Y 兲 ja ␴ H ⫹ji ␴ 关 X⫹Y 兴 S. Dalgarno, Proc. R. Soc. London, Ser. A 247, 243 共1958兲.
28
Q ai ␴ ⫽ 29
N. C. Handy and H. F. Schaefer, J. Chem. Phys. 81, 5031 共1984兲.
30
T. Helgaker and P. Jørgensen, Theor. Chim. Acta 75, 111 共1989兲.
31
⫹ 共 X⫺Y 兲 ia ␴ H ⫺
ji ␴ 关 X⫺Y 兴 其 , 共A9兲 O. Christiansen, P. Jørgensen, and C. Hättig, Int. J. Quantum Chem. 68, 1
共1998兲.
and for the virt–virt block, with a⭐b,
32
A. E. Hansen and T. D. Bouman, J. Am. Chem. Soc. 107, 4828 共1985兲.
33
H. Weiss, R. Ahlrichs, and M. Häser, J. Chem. Phys. 99, 1262 共1993兲.
Q ab ␴ ⫽W ab ␴ , 共A10兲 34
A. Görling, H. H. Heinze, S. P. Ruzankin, M. Stamfer, and N. Rösch, J.
Chem. Phys. 110, 2785 共1999兲.
35
S. J. A. van Gisbergen, J. G. Snijders, and E. J. Baerends, Comput. Phys.
Q ab ␴ ⫽ 兺i ⍀ 兵 共 X⫹Y 兲 ia ␴共 X⫺Y 兲 ib ␴ ⫹ 共 X⫺Y 兲 ia ␴共 X⫹Y 兲 ib ␴ 其 36
Commun. 118, 119 共1999兲.
O. Treutler and R. Ahlrichs, J. Chem. Phys. 102, 346 共1995兲.
37
R. E. Stratmann, G. E. Scuseria, and M. J. Frisch, Chem. Phys. Lett. 257,
⫹ 兺i ⑀ i ␴ 兵 共 X⫹Y 兲 ia ␴共 X⫹Y 兲 ib ␴ 38
213 共1996兲.
B. G. Johnson, P. M. W. Gill, and J. A. Pople, J. Chem. Phys. 98, 5612
共5612兲.
39
⫹ 共 X⫺Y 兲 ia ␴ 共 X⫺Y 兲 ib ␴ 其 . 共A11兲 R. Ahlrichs, M. Bär, M. Häser, H. Horn, and C. Kölmel, Chem. Phys. Lett.
162, 165 共1989兲, current version: see http://www.turbomole.de
Subtracting Eq. 共A8兲 from 共A6兲, we obtain 40
K. Eichkorn, O. Treutler, H. Öhm, M. Häser, and R. Ahlrichs, Chem.
Phys. Lett. 242, 652 共1995兲.
⫹
共 ⑀ a ␴ ⫺ ⑀ i ␴ 兲 Z ia ␴ ⫹H ia ␴ 关 Z 兴 ⫽⫺ 共 Q ia ␴ ⫺Q ai ␴ 兲 , 共A12兲 41
R. Bauernschmitt, M. Häser, O. Treutler, and R. Ahlrichs, Chem. Phys.
Lett. 264, 573 共1997兲.
which is the Z vector equation 共18兲 with the RHS given in 42
A. Schäfer, H. Horn, and R. Ahlrichs, J. Chem. Phys. 97, 2571 共1992兲.
Eq. 共21兲. Once the Z vector equation has been solved, W is 43
A. Schäfer, C. Huber, and R. Ahlrichs, J. Chem. Phys. 100, 5829 共1994兲.
44
determined from Eqs. 共A4兲, 共A10兲, and the sum of Eqs. 共A8兲 R. A. Kendall, T. H. Dunning, Jr., and R. J. Harrison, J. Chem. Phys. 96,
6796 共1992兲.
and 共A6兲 according to 45
T. H. Dunning, Jr., J. Chem. Phys. 90, 1007 共1989兲.
46
D. E. Woon and T. H. Dunning, Jr., J. Chem. Phys. 98, 1358 共1993兲.
W i j ␴ ⫽Q i j ␴ ⫹H ⫹
i j␴关 Z 兴 , 47
J. Čı́z̃ek and J. Paldus, J. Chem. Phys. 47, 3976 共1967兲.
48
R. Bauernschmitt and R. Ahlrichs, J. Chem. Phys. 104, 9047 共1996兲.
W ab ␴ ⫽Q ab ␴ , 共A13兲 49
M. E. Casida, C. Jamorski, K. C. Casida, and D. R. Salahub, J. Chem.
Phys. 108, 4439 共1998兲.
W ia ␴ ⫽Q ai ␴ ⫹ ⑀ i ␴ Z ia ␴ . 50
D. J. Tozer and N. C. Handy, J. Chem. Phys. 109, 10180 共1998兲.
51
This is identical to the definition of W as given in Eq. 共24兲. C. Adamo, G. E. Scuseria, and V. Barone, J. Chem. Phys. 111, 2889
共1999兲.
52
P. A. M. Dirac, Proc. R. Soc. London, Ser. A 123, 714 共1929兲.
1
Structure and Dynamics of Electronic Excited States, edited by J. Laane
53
J. C. Slater, Phys. Rev. 81, 385 共1951兲.
共Springer, Berlin, 1999兲.
54
S. H. Vosko, L. Wilk, and M. Nusair, Can. J. Phys. 58, 1200 共1980兲.
2
M. Klessinger and J. Michl, Excited States and Photochemistry of Organic
55
J. P. Perdew and Y. Wang, Phys. Rev. B 45, 13244 共1992兲.
Molecules 共VCH, Weinheim, 1995兲.
56
A. D. Becke, Phys. Rev. A 38, 3098 共1988兲.
3
W. Domcke and G. Stock, Adv. Chem. Phys. 100, 1 共1997兲.
57
C. Lee, W. Yang, and R. G. Parr, Phys. Rev. B 37, 785 共1988兲.
4
J. M. Hollas, High Resolution Spectroscopy 共Wiley, New York, 1998兲.
58
J. P. Perdew, Phys. Rev. B 33, 8822 共1986兲.
5 59
J. B. Foresman, M. Head-Gordon, J. A. Pople, and M. J. Frisch, J. Phys. J. P. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. 77, 3865
Chem. 96, 135 共1992兲. 共1996兲.
60
6
D. Maurice and M. Head-Gordon, Mol. Phys. 96, 1533 共1999兲. J. P. Perdew, M. Ernzerhof, and K. Burke, J. Chem. Phys. 105, 9982
7
J. F. Stanton, J. Gauss, N. Ishikawa, and M. Head-Gordon, J. Chem. Phys. 共1996兲.
103, 4160 共1995兲.
61
J. P. Perdew and A. Zunger, Phys. Rev. B 23, 5048 共1981兲.
8 62
M. E. Casida, in Recent Advances in Computational Chemistry, edited by The TURBOMOLE basis set library is available via anonymous ftp from
D. P. Chong 共World Scientific, Singapore, 1995兲, Vol. 1, p. 155. ftp://ftp.chemie.uni-karlsruhe.de/pub/basen
63
9
R. Bauernschmitt and R. Ahlrichs, Chem. Phys. Lett. 256, 454 共1996兲. R. Krishnan, J. S. Binkley, R. Seeger, and J. A. Pople, J. Chem. Phys. 72,
10
R. E. Stratmann, G. E. Scuseria, and M. J. Frisch, J. Chem. Phys. 109, 650 共1980兲.
8218 共1998兲. 64
K. P. Huber and G. Herzberg, Molecular Spectra and Molecular Structure
11
S. Hirata and M. Head-Gordon, Chem. Phys. Lett. 314, 291 共1999兲. 共Van Nostrand Reinhold, New York, 1979兲, Vol. IV.
12
C. Van Caillie and R. D. Amos, Chem. Phys. Lett. 308, 249 共1999兲. 65
G. Herzberg, Molecular Spectra and Molecular Structure 共Van Nostrand
13
C. Van Caillie and R. D. Amos, Chem. Phys. Lett. 317, 159 共2000兲. Reinhold, New York, 1966兲, Vol. II.
14
A. J. Sobolewski and W. Domcke, Phys. Chem. Chem. Phys. 1, 3065 66
C. W. Bauschlicher and S. R. Langhoff, J. Chem. Phys. 85, 5936 共1986兲.
共1999兲. 67
D. J. Clouthier and J. Karolczak, J. Chem. Phys. 94, 1 共1991兲.

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions
 J. Chem. Phys., Vol. 117, No. 16, 22 October 2002 TDDFT for excited state properties 7447

68
D. J. Clouthier and D. A. Ramsay, Annu. Rev. Phys. Chem. 34, 31 共1983兲. 72
J. M. Hollas, in Recent Experimental and Computational Advances in
69
O. Christiansen, J. F. Stanton, and J. Gauss, J. Chem. Phys. 108, 3987 Molecular Spectroscopy, edited by R. Fausto 共Kluwer, Dordrecht, 1993兲,
共1998兲. p. 27.
70 73
A. Bolovinos, P. Tsekeris, J. Philis, E. Pantos, and G. Andritsopoulos, J. M. Drabbels, W. L. Meerts, and J. J. ter Meulen, J. Chem. Phys. 99, 2352
Mol. Spectrosc. 103, 240 共1984兲. 共1993兲.
71 74
L. Edwards, D. H. Dolphin, M. Gouterman, and A. D. Adler, J. Mol. R. Y. Dong, R. Nanes, and D. A. Ramsay, Can. J. Chem. 71, 1595
Spectrosc. 38, 16 共1971兲. 共1993兲.

Downloaded 07 Apr 2013 to 113.128.225.91. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://jcp.aip.org/about/rights_and_permissions